698 Background: Randomized trials of first-line single-agent checkpoint inhibitors (CPIs) in advanced urothelial cancer (aUC) have shown no overall survival benefit compared to chemotherapy. However, a subset of patients may derive long-term benefit from CPIs. Identifying these patients is challenging due to their delayed treatment effects. We investigated whether machine learning (ML)-based phenotyping within a trial emulation framework could identify patients with improved survival from pembrolizumab vs gemcitabine/carboplatin (GCa) in a patient population similar to KEYNOTE-361. Methods: We used the Flatiron Health nationwide, de-identified EHR-derived database. Patients receiving first-line pembrolizumab or GCa were included. A supervised gradient-boosted survival model using pre-treatment variables (demographics, cancer features, biomarkers, labs, ICD codes) was developed to predict early mortality. Patients in the highest-risk tertile were excluded from the emulated trial, as they were unlikely to benefit from CPIs’ delayed treatment effect. Survival was compared using restricted mean survival time (RMST) and survival probabilities at 3 and 4 years from inverse probability–weighted Kaplan–Meier curves. Results: Among 3706 patients (pembrolizumab, n = 1484; GCa, n = 2222), early mortality was higher with pembrolizumab, but 3-year survival was modestly improved (25.2% vs 19.1%; Δ6.1 95% CI 2.4-9.8). RMST was similar at 3 years (15.9 vs 15.9 months; Δ0.0 95% CI -1.0-1.1) and 4 years (18.6 vs 17.9 months; Δ0.6 95% CI -0.7-2.0). The ML model had strong discriminatory performance (AUCs of 0.78 at 90 days and 0.72 at 1 year). After excluding the highest-risk tertile, pembrolizumab showed superior 3-year survival (33.1% vs 24.7%; Δ8.5 95% CI 3.7-13.5) and lower early mortality. RMST differences showed a trend toward improvement at 3 years (19.1 vs 18.1 months; Δ1.0 95% CI -0.3-2.3) and favored pembrolizumab at 4 years (22.6 vs 20.8 months; Δ1.8 95% CI 0.1-3.6). PD-L1 testing was infrequent (< 10%) and not predictive of survival outcomes. Conclusions: ML-based phenotyping within a trial emulation framework identified a subgroup of aUC patients with improved long-term survival with pembrolizumab monotherapy. While absolute gains were modest, this data-driven approach may help refine patient selection for CPIs with heterogeneous treatment effects.
Orcutt et al. (Sun,) studied this question.